Fairness risk measures

Robert C. Williamson, Aditya Krishna Menon

Ensuring that classifiers are non-discriminatory with respect to sensitive features (like race or gender) is a critical challenge in modern machine learning. This paper bridges the gap between machine learning fairness and mathematical finance by reframing fairness as a "risk management" problem. The authors propose a new, convex framework that uses the Conditional Value at Risk (CVaR) to minimize performance disparities across different subgroups, providing a flexible way to handle complex sensitive features while maintaining a reliable optimization landscape.

The Problem with Current Fairness Definitions

Machine learning models often inadvertently discriminate against specific groups. While many "fairness" definitions exist, they frequently suffer from two major flaws:

  1. Narrow Scope: They often assume sensitive features are strictly binary (e.g., "Group A" vs. "Group B"). This makes them difficult to apply to categorical or real-valued data where multiple intersecting subgroups exist.
  2. Mathematical Complexity: Many fairness constraints result in "non-convex" optimization problems. For engineers, this is a nightmare—it means the training process can easily get stuck in local minima, making models unstable, difficult to tune, and computationally expensive to train.

The authors seek a definition of fairness that is mathematically "well-behaved" (convex) and general enough to work regardless of the number or type of subgroups.

Methodology: Fairness as Risk Management

The core idea is to move away from looking at a single global "risk" (average loss) and instead look at a distribution of subgroup risks.

  • Subgroup Risk Aggregation: The authors define the risk for each subgroup induced by a sensitive feature. They then propose aggregating these risks using a functional that measures both the mean and the deviation (disparity) between groups.
  • Axiomatic Framework: Rather than picking an arbitrary formula, they define seven mathematical axioms (such as Convexity, Monotonicity, and Aversity) that any truly "fair" aggregator should satisfy.
  • The Finance Connection: They prove that any measure satisfying these axioms is mathematically equivalent to a "coherent and regular measure of risk"—the exact same tools used in mathematical finance to manage portfolio volatility and downside risk.
  • CVaR-Fairness: As a practical application, they focus on Conditional Value at Risk (CVaR). In this context, CVaR focuses on the "tail" of the subgroup risk distribution. Essentially, it forces the model to improve performance for the specific groups that are currently faring the worst.

Key Insights

  • The Philosophical Spectrum: The CVaR objective includes a tuning parameter, $\alpha$.
  • When $\alpha \to 0$, the model acts like a standard "utilitarian" learner, caring only about average overall performance.
  • When $\alpha \to 1$, it adopts a "Rawlsian Maximin" approach, focusing entirely on lifting up the group with the worst performance.
  • This allows practitioners to choose exactly where they want to sit on the spectrum of social justice versus pure utility.
  • Convexity is King: By framing fairness through these specific risk measures, the resulting objective function remains convex (provided the base loss is convex). This ensures that adding fairness doesn't break the efficiency and reliability of standard gradient-based training.

Key Results

The authors evaluated their approach on two benchmarks: synth (a synthetic two-dimensional dataset designed to test fairness-accuracy tradeoffs) and the UCI adult dataset (a real-world census dataset predicting income, using gender as the sensitive attribute).

Tunable Tradeoffs and Performance On both datasets, varying the $\alpha$ parameter allowed for a smooth transition between high accuracy (with high disparity) and high fairness (with lower accuracy). The CVaR-based approach achieved fairness-accuracy tradeoffs that were competitive with or superior to existing state-of-the-art baselines like "Fair-ERM," while being more mathematically robust.

Fairness-accuracy tradeoff comparing CVaR, SVM, and Fair-ERM.

Consistency in Leveling the Playing Field As the tuning parameter $\alpha$ increased toward 1, the absolute difference in risk between subgroups consistently dropped. This confirms that the model successfully "levels the playing field" by prioritizing the most disadvantaged groups.

Subgroup risk difference as a function of the tuning parameter α.

Limitations and Future Work

  • Surrogate vs. Real Loss: The optimization uses "surrogate" losses (like square-hinge loss) because they are differentiable. However, practitioners often care about the 0-1 error (accuracy). While the method works well on surrogates, the direct relationship to 0-1 fairness requires further study.
  • Conditioning: The main framework focuses on "Demographic Parity." While the authors touch on "Equality of Opportunity" (conditioning on the true label) in the appendix, more work is needed to fully integrate complex conditional fairness notions into the risk measure framework.
  • Beyond CVaR: CVaR is just one type of risk measure. The authors suggest that other financial tools, like spectral measures or entropic value at risk, might offer even better properties for specific fairness applications in the future.

Summary updated

Abstract

Ensuring that classifiers are non-discriminatory or fair with respect to a sensitive feature (e.g., race or gender) is a topical problem. Progress in this task requires fixing a definition of fairn...
BibTeX
@inproceedings{williamson2019fairness,
  author = {Robert C. Williamson and Aditya Krishna Menon},
  title = {{Fairness risk measures}},
  booktitle = {International Conference on Machine Learning},
  year = {2019},
  url = {https://openreview.net/forum?id=B1-_OsZOWB},
}
Figure